How it works
What a visitor actually watches happen at kineticml.com, grounded in the shipping code, not the brochure version. Enter a business type and a city, or paste a website directly. Pick one of the matches. Pass a short capture gate (name, email, company, two free runs). Then watch the research happen live, one step at a time, with a timeline showing what is pending, active and done. It ends with a named decision maker, a discovered email with a confidence score, and outreach drafts written from what the pipeline just found.
The eight steps: business details (name, address, phone, rating, review count and website pulled via Google Places); a website scan (headline, about text, services and contact emails, flagging booking, live chat and a review widget, the gaps the outreach will reference); Google reviews (rating, review count, and quotable excerpts with author names); social discovery (Instagram, Facebook, TikTok and LinkedIn, honest about platforms checked but not found); decision maker identification (a seven-source waterfall: website team page, Google search, Yelp, BBB, LinkedIn search, review-text mining and name parsing from contact emails, returning the owner or manager with title and source, or stating plainly that no one was identified); personal profiles (the decision maker's personal LinkedIn, Instagram, Facebook and TikTok); email discovery (website contact emails first, then nine name-and-domain patterns, then generic addresses, with a primary email and an explicit confidence score); and writing the outreach (personalized copy with dynamic subject lines referencing specifics the pipeline just surfaced, Claude writing first with a fallback model and a template engine behind it so the demo always completes).
The pipeline reports its misses on screen: "checked, not found" for socials, "not identified" with sources listed for decision makers, and confidence percentages on every email. That honesty is deliberate; a demo that never misses reads as staged, one that shows its work earns trust.
The engine is built for the thin end of the market: local and small businesses whose owners are not in any B2B contact database. Every visitor who passes the capture gate lands in the client's CRM tagged as demo traffic, with the business they researched attached as context, and the call to action is a booked conversation about the client's full outreach service. The funnel is the tool.
The technical shape: a Next.js, React and TypeScript front end on Tailwind, deployed on Vercel, with Framer Motion rendering the live-pipeline timeline as each step's event arrives; a FastAPI backend on Python, Dockerized and deployed on Railway (v3.1.0), running async scraping with server-sent events streaming one event per step, and endpoints for search, enrichment by place, enrichment by direct URL, and lead capture. Google Places handles business search and details; Anthropic Claude is the primary writer for outreach emails, with OpenAI as a fallback writer and a template engine behind both; research runs on public-web scraping and Google-indexed reads of Yelp, BBB, LinkedIn and the major social platforms, with all data sourced from public records and websites, as the product's own footer states.
Scope honesty: discovered emails are confidence-scored pattern work, not verified deliverability, and the score is shown rather than hidden. The demo neither sends email nor stores a visitor's research beyond the session; the engine is industry-agnostic, with the interface's examples as placeholders rather than a vertical limit.
The market it sits in
Sales intelligence is crowded at the top of the market and thin at the bottom. Kinetic ML works the thin end, where the decision maker is found by triangulation rather than looked up in a database. Contact databases such as ZoomInfo and Apollo sell pre-built B2B contact records, strongest on funded, LinkedIn-dense companies; enrichment orchestration such as Clay sells a workbench that chains data providers together; email finding tools such as Hunter.io sell pattern discovery as a point solution; the current wave of AI SDRs sells automated sequencing on top of purchased data. Kinetic ML runs the research in front of the prospect, on a business they name, and shows its misses along with its hits; the demo is the differentiation, not a feature list.
For the record: the client's product is "Kinetic ML," two words, at kineticml.com. The kineticai.io domain Common Ground holds is a separate, unrelated property.
If it becomes ongoing work
This was a contract build, not a Common Ground venture. If the client ever wants the engine maintained and extended as ongoing infrastructure rather than a one-time delivery, that is a future lane Robert Prochnow would run: keeping the scraper waterfall healthy as platforms drift, and wiring it into the client's own CRM and sequencing over time.
What it produced
Built and delivered in 2025, live in public at kineticml.com, with the backend matured to v3.1.0 in early 2026. Every completed run lands a captured lead in the client's CRM. The engine is running the client's pitch today.
The result, in short
Built and delivered in 2025, live in public at kineticml.com, with the backend matured to v3.1.0 in early 2026. Every completed run lands a captured lead in the client's CRM. The engine is live and running the client's outreach pitch today.
A slice of the project list
A few related projects.
- Continuum: Common Ground's internal lab for the digital-twin line
- Sublime Medical: fractional COO work for a California cosmetic dermatology group, tracing a collections problem to stale billing codes (2015)
- Soft-Story Retrofits: launched a retrofit and PACE financing line against Los Angeles's mandatory seismic ordinance (2017)